基于Spark-0.4和Hadoop-0.20.2

1. Kmeans

数据：自己产生的三维数据，分别围绕正方形的8个顶点

{0, 0, 0}, {0, 10, 0}, {0, 0, 10}, {0, 10, 10},

{10, 0, 0}, {10, 0, 10}, {10, 10, 0}, {10, 10, 10}

Point number

189,918,082 （1亿9千万个三维点）

Capacity

10GB

HDFS Location

/user/LijieXu/Kmeans/Square-10GB.txt

Hadoop vs Spark性能对比_hdfs

程序逻辑：

读取HDFS上的block到内存，每个block转化为RDD，里面包含vector。

然后对RDD进行map操作，抽取每个vector（point）对应的类号，输出（K,V）为（class，（Point，1）），组成新的RDD。

然后再reduce之前，对每个新的RDD进行combine，在RDD内部算出每个class的中心和。使得每个RDD的输出只有最多K个KV对。

最后进行reduce得到新的RDD（内容的Key是class，Value是中心和，再经过map后得到最后的中心。

先上传到HDFS上，然后在Master上运行

root@master:/opt/spark# ./run spark.examples.SparkKMeans master@master:5050 hdfs://master:9000/user/LijieXu/Kmeans/Square-10GB.txt 8 2.0

迭代执行Kmeans算法。

一共160个task。（160 * 64MB = 10GB）

利用了32个CPU cores，18.9GB的内存。

每个机器的内存消耗为4.5GB （共40GB）（本身points数据10GB*2，Map后中间数据(K, V) => (int, (vector, 1)) （大概10GB）

Hadoop vs Spark性能对比_数据_02

最后结果：

0.505246194 s
Final centers: Map(5 -> (13.997101228817169, 9.208875044622895, -2.494072457488311), 8 -> (-2.33522333047955, 9.128892414676326, 1.7923150585737604), 7 -> (8.658031587043952, 2.162306996983008, 17.670646829079146), 3 -> (11.530154433698268, 0.17834347219956842, 9.224352885937776), 4 -> (12.722903153986868, 8.812883284216143, 0.6564509961064319), 1 -> (6.458644369071984, 11.345681702383024, 7.041924994173552), 6 -> (12.887793408866614, -1.5189406469928937, 9.526393664105957), 2 -> (2.3345459304412164, 2.0173098597285533, 1.4772489989976143))

50MB/s 10GB => 3.5min

10MB/s 10GB => 15min

在20GB的数据上测试

Point number

377,370,313 （3亿7千万个三维点）

Capacity

20GB

HDFS Location

/user/LijieXu/Kmeans/Square-20GB.txt

运行测试命令：

root@master:/opt/spark# ./run spark.examples.SparkKMeans master@master:5050 hdfs://master:9000/user/LijieXu/Kmeans/Square-20GB.txt 8 2.0 | tee mylogs/sqaure-20GB-kmeans.log

得到聚类结果：

Final centers: Map(5 -> (-0.47785701742763115, -1.5901830956323306, -0.18453046159033773), 8 -> (1.1073911553593858, 9.051671594514225, -0.44722211311446924), 7 -> (1.4960397239284795, 10.173412443492643, -1.7932911100570954), 3 -> (-1.4771114031182642, 9.046878176063172, -2.4747981387714444), 4 -> (-0.2796747780312184, 0.06910629855122015, 10.268115903887612), 1 -> (10.467618592186486, -1.168580362309453, -1.0462842137817263), 6 -> (0.7569895433952736, 0.8615441990490469, 9.552726007309518), 2 -> (10.807948500515304, -0.5368803187391366, 0.04258123037074164))

基本就是8个中心点

内存消耗：（每个节点大约5.8GB），共50GB左右。

Hadoop vs Spark性能对比_Spark_03

内存分析：

20GB原始数据，20GB的Map输出

迭代次数

时间

1

108 s

2

0.93 s

12/06/05 11:11:08 INFO spark.CacheTracker: Looking for RDD partition 2:302

12/06/05 11:11:08 INFO spark.CacheTracker: Found partition in cache!

在20GB的数据上测试（迭代更多的次数）

root@master:/opt/spark# ./run spark.examples.SparkKMeans master@master:5050 hdfs://master:900
0/user/LijieXu/Kmeans/Square-20GB.txt 8 0.8

Task数目：320

时间：

迭代次数

时间

1

100.9 s

2

0.93 s

3

4.6 s

4

3.9 s

5

3.9 s

6

3.9 s

迭代轮数对内存容量的影响：

Hadoop vs Spark性能对比_spark_04

基本没有什么影响，主要内存消耗：20GB的输入数据RDD，20GB的中间数据。

Final centers: Map(5 -> (-4.728089224526789E-5, 3.17334874733142E-5, -2.0605806380414582E-4), 8 -> (1.1841686358289191E-4, 10.000062966002101, 9.999933240005394), 7 -> (9.999976672588097, 10.000199556926772, -2.0695123602840933E-4), 3 -> (-1.3506815993198176E-4, 9.999948270638338, 2.328148782609023E-5), 4 -> (3.2493629851483764E-4, -7.892413981250518E-5, 10.00002515017671), 1 -> (10.00004313126956, 7.431996896171192E-6, 7.590402882208648E-5), 6 -> (9.999982611661382, 10.000144597573051, 10.000037734639696), 2 -> (9.999958673426654, -1.1917651103354863E-4, 9.99990217533504))

结果可视化

Hadoop vs Spark性能对比_Spark_05

2. HdfsTest

测试逻辑：

package spark.examples
import spark._
object HdfsTest {
def main(args: Array[String]) {
val sc = new SparkContext(args(0), "HdfsTest")
val file = sc.textFile(args(1))
val mapped = file.map(s => s.length).cache()
for (iter <- 1 to 10) {
val start = System.currentTimeMillis()
for (x <- mapped) { x + 2 }
// println("Processing: " + x)
val end = System.currentTimeMillis()
println("Iteration " + iter + " took " + (end-start) + " ms")
}
}
}

首先去HDFS上读取一个文本文件保存在file

再次计算file中每行的字符数，保存在内存RDD的mapped中

然后读取mapped中的每一个字符数，将其加2，计算读取+相加的耗时

只有map，没有reduce。

测试10GB的Wiki

实际测试的是RDD的读取性能。


root@master:/opt/spark# ./run spark.examples.HdfsTest master@master:5050 hdfs://master:9000:/user/LijieXu/Wikipedia/txt/enwiki-20110405.txt

测试结果：

Iteration 1 took 12900 ms = 12s
Iteration 2 took 388 ms
Iteration 3 took 472 ms
Iteration 4 took 490 ms
Iteration 5 took 459 ms
Iteration 6 took 492 ms
Iteration 7 took 480 ms
Iteration 8 took 501 ms
Iteration 9 took 479 ms
Iteration 10 took 432 ms

每个node的内存消耗为2.7GB （共9.4GB * 3）

Hadoop vs Spark性能对比_spark_06

实际测试的是RDD的读取性能。

root@master:/opt/spark# ./run spark.examples.HdfsTest master@master:5050 hdfs://master:9000/user/LijieXu/Wikipedia/txt/enwiki-20110405.txt
测试90GB的RandomText数据
root@master:/opt/spark# ./run spark.examples.HdfsTest master@master:5050 hdfs://master:9000/user/LijieXu/RandomText90GB/RandomText90GB

耗时：

迭代次数

耗时

1

111.905310882 s

2

4.681715228 s

3

4.469296148 s

4

4.441203887 s

5

1.999792125 s

6

2.151376037 s

7

1.889345699 s

8

1.847487668 s

9

1.827241743 s

10

1.747547323 s

内存总消耗30GB左右。

单个节点的资源消耗：

Hadoop vs Spark性能对比_hdfs_07

3. 测试WordCount

写程序：

import spark.SparkContext
import SparkContext._
object WordCount {
def main(args: Array[String]) {
if (args.length < 2) {
System.err.println("Usage: wordcount <master> <jar>")
System.exit(1)
}
val sp = new SparkContext(args(0), "wordcount", "/opt/spark", List(args(1)))
val file = sp.textFile("hdfs://master:9000/user/LijieXu/Wikipedia/txt/enwiki-20110405.txt");
val counts = file.flatMap(line => line.split(" ")).map(word => (word, 1)).reduceByKey(_ + _)
counts.saveAsTextFile("hdfs://master:9000/user/Output/WikiResult3")
}
}

打包成mySpark.jar，上传到Master的/opt/spark/newProgram。

运行程序：

root@master:/opt/spark# ./run -cp newProgram/mySpark.jar WordCount master@master:5050 newProgram/mySpark.jar

Mesos自动将jar拷贝到执行节点，然后执行。

内存消耗：（10GB输入file + 10GB的flatMap + 15GB的Map中间结果（word，1））

还有部分内存不知道分配到哪里了。

耗时：50 sec（未经过排序）

Hadoop WordCount耗时：120 sec到140 sec

结果未排序

单个节点：

Hadoop vs Spark性能对比_Spark_08

Hadoop测试

Kmeans

运行Mahout里的Kmeans

root@master:/opt/mahout-distribution-0.6# bin/mahout org.apache.mahout.clustering.syntheticcontrol.kmeans.Job -Dmapred.reduce.tasks=36 -i /user/LijieXu/Kmeans/Square-20GB.txt -o output -t1 3 -t2 1.5 -cd 0.8 -k 8 -x 6

在运行（320个map，1个reduce）

Canopy Driver running buildClusters over input: output/data

时某个slave的资源消耗情况

Hadoop vs Spark性能对比_hdfs_09

Hadoop vs Spark性能对比_hdfs_10

Completed Jobs

Jobid

Name

Map Total

Reduce Total

Time

job_201206050916_0029

Input Driver running over input: /user/LijieXu/Kmeans/Square-10GB.txt

160

0

1分2秒

job_201206050916_0030

KMeans Driver running runIteration over clustersIn: output/clusters-0/part-randomSeed

160

1

1分6秒

job_201206050916_0031

KMeans Driver running runIteration over clustersIn: output/clusters-1

160

1

1分7秒

job_201206050916_0032

KMeans Driver running runIteration over clustersIn: output/clusters-2

160

1

1分7秒

job_201206050916_0033

KMeans Driver running runIteration over clustersIn: output/clusters-3

160

1

1分6秒

job_201206050916_0034

KMeans Driver running runIteration over clustersIn: output/clusters-4

160

1

1分6秒

job_201206050916_0035

KMeans Driver running runIteration over clustersIn: output/clusters-5

160

1

1分5秒

job_201206050916_0036

KMeans Driver running clusterData over input: output/data

160

0

55秒

job_201206050916_0037

Input Driver running over input: /user/LijieXu/Kmeans/Square-20GB.txt

320

0

1分31秒

job_201206050916_0038

KMeans Driver running runIteration over clustersIn: output/clusters-0/part-randomSeed

320

36

1分46秒

job_201206050916_0039

KMeans Driver running runIteration over clustersIn: output/clusters-1

320

36

1分46秒

job_201206050916_0040

KMeans Driver running runIteration over clustersIn: output/clusters-2

320

36

1分46秒

job_201206050916_0041

KMeans Driver running runIteration over clustersIn: output/clusters-3

320

36

1分47秒

job_201206050916_0042

KMeans Driver running clusterData over input: output/data

320

0

1分34秒

运行多次10GB、20GB上的Kmeans，资源消耗

Hadoop vs Spark性能对比_hdfs_11

Hadoop vs Spark性能对比_hdfs_12

Hadoop WordCount测试

Hadoop vs Spark性能对比_Spark_13

Hadoop vs Spark性能对比_Spark_14

Spark交互式运行

进入Master的/opt/spark

运行

MASTER=master@master:5050 ./spark-shell

打开Mesos版本的spark

在master:8080可以看到framework

Active Frameworks

ID

User

Name

Running Tasks

CPUs

MEM

Max Share

Connected

201206050924-0-0018

root

Spark shell

0

0

0.0 MB

0.00

2012-06-06 21:12:56

scala> val file = sc.textFile("hdfs://master:9000/user/LijieXu/Wikipedia/txt/enwiki-20110405.txt")
scala> file.first
scala> val words = file.map(_.split(' ')).filter(_.size < 100) //得到RDD[Array[String]]
scala> words.cache
scala> words.filter(_.contains("Beijing")).count
12/06/06 22:12:33 INFO SparkContext: Job finished in 10.862765819 s
res1: Long = 855
scala> words.filter(_.contains("Beijing")).count
12/06/06 22:12:52 INFO SparkContext: Job finished in 0.71051464 s
res2: Long = 855
scala> words.filter(_.contains("Shanghai")).count
12/06/06 22:13:23 INFO SparkContext: Job finished in 0.667734427 s
res3: Long = 614
scala> words.filter(_.contains("Guangzhou")).count
12/06/06 22:13:42 INFO SparkContext: Job finished in 0.800617719 s
res4: Long = 134

由于GC的问题，不能cache很大的数据集。